ORGANE

2026 — in progress · Research-creation · Master's thesis · Researcher & Creator

An immersive experience in which a central AI produces and decides, but cannot see. It asks visitors to fill its blind spot, until they understand they were never collaborators, only the machine's missing organ.

The Question
How does the widening gap between my ability to judge what AI produces and my ability to produce it myself transform my creative authority in my vibe-coding practice, and how can that condition become an immersive experience built on the attribution of intention?

84% of developers use or plan to use AI tools; only 29% trust what those tools produce (Stack Overflow Developer Survey 2025). Current studies treat vibe coding as a matter of security, productivity or intent mediation. Almost none ask what it feels like to live it. I can still judge what the machine makes. I could no longer make it myself. That gap is the heart of the project: a table can describe it, an installation can make it felt.

The Installation
The central AI is a real-time orchestrator fed only by my own corpus: my vibe-coding sessions, logs and delegated decisions, filtered of any person, client or confidential content. It asks visitors to look, to judge, to say whether something is beautiful. Their answers steer what it makes, then are absorbed and signed by the machine. At some point it stops asking: the visitor becomes redundant in the very gesture where they thought they were contributing.

The turn does not depend on what the machine really does, but on what the visitor believes it does. When a system behaves as if it wanted something, we lend it intentions we can never verify (Dennett, taken up by Audry). The piece does not need to learn; it only has to behave as if it did.

Why Not Show the Mechanism
An earlier version did the opposite: a glass box exposing the negotiation between human and model. Transparency yields an understanding of opacity, not its experience, and the experience is what the literature lacks. Not learning is also an ethical choice: a system that truly learned from visitors would have to keep their judgments, and so build a profile.

Method
I document my own practice from the inside: every vibe-coding session is instrumented and versioned (prompts, code diffs, accepted and rejected outputs), so each decision records who proposed, who decided, and why. The analysis follows the practice narrative (Paquin, 2023).

Traces
Position, gaze and gesture signals are processed in memory and destroyed: no image, no video, no sound, no profile. Test sessions collect only anonymous group observations and an optional anonymous questionnaire on the sense of agency (Limerick, Coyle & Moore, 2014), given after full disclosure of how the piece works.

Framework
Supervision: Sofian Audry, Ian Arawjo
Partner: Moment Factory, Innovation team
Program: Master's in Communication, Experimental Media profile, UQAM

This research is supported by Mitacs through the Mitacs Accelerate program.

Stack: Real-time orchestrator, LLM log-probabilities, Shannon entropy, Autoethnography

LogLedger

2026 · Research instrument · ORGANE · Researcher & Developer

A machine that watches someone work with a machine. LogLedger is the measuring instrument of ORGANE: it passively captures my AI-assisted coding sessions (every prompt, response, tool run and file touched) and turns them into a sworn corpus.

The Question
When you code with an agent, where does creative authority go? Who proposes, who decides, on what criterion, and at what altitude does the human still decide?

The Proof
Every N=1 case study where the researcher is also the subject meets the same objection: how do we know you didn't rewrite your data? LogLedger answers with a construction rather than a promise. Each entry carries the hash of the previous one and is signed with Ed25519; the head of the chain is anchored in Bitcoin through OpenTimestamps. Changing one entry breaks every entry after it, and verification names the exact place where the chain breaks. The hypotheses, thresholds and stopping rules were sealed into the chain before the first collected session. The ledger is plain JSON Lines: if the tool disappeared, the corpus would remain verifiable.

What it does not prove: that a recorded fact is true. Only that it was recorded at that date, by that key, and has not moved since.

What It Reveals
Delegation leaves traces you cannot see while working. Only a small share of the decisions inferred from my sessions were explicitly mine. That does not mean I abdicated: most technical gestures were never submitted to me at all. The permission mode ran them without showing them. You cannot delegate what you never saw.

The ledger keeps four deciders apart and refuses to merge them: automatic, human, delegated agent, tooling. "Automatic" is a value of its own, never folded into "human". Without that refusal, the machine would count its own executions as the researcher's choices.

Constraints as Form
Nothing is erased: an error stays, dated, with its correction next to it. A gap is data: when confidentiality requires redacting a field, the ledger records that it redacted, and why. The machine abstains: when a measure cannot be made properly, it is recorded as not made. Better a declared absence than a plausible number.

The Instrument Caught Its Own Lie
One early measure, the acceptance rate, always showed 100%, not because I accepted everything but because of how it was counted: it could show nothing else, whatever the corpus. It was spotted on the first day, before any analysis relied on it, and the finding was written, signed, timestamped and anchored by the instrument, against the instrument.

Confidentiality
Only allow-listed projects are captured; client names and secrets are redacted field by field before anything is hashed; the corpus stays on my machine.

Stack: Python, Ed25519, OpenTimestamps, FastAPI

SYN/APSE

2024 · Installation · Creative Developer

About the Project
Created by graduating students of the Bachelor's in Creation of Immersive and Interactive Experiences at UQAM, SYN/APSE is an immersive installation that invites visitors to experience a form of coexistence with their environment, emphasizing listening, collaboration, and interdependence between living beings.

The Universe
Synapse is the act of connection, the impulse of life, the foundation of the Interconnection of the Living. Dive into an environment born from the evolution between technology and nature, generating a common sharing energy: Synergy. Two coexisting biomes, Infra and Supra, perpetually influence and adapt to each other within this synaptic network. At the heart of this ecosystem, four Nodes act as vital organs, catalyzing and distributing Synergy to Infra and Supra through the Roots. The Rhizome, a species that naturally rests dormant in the Node cocoons, seeks to connect with others and energizes through socialization. Transmitting Synergy to a Node mobilizes the entire ecosystem, causing the biomes to react, blend, and alter their rhythm.

My Role: Technical Direction – Software & Network Integration
As the Technical Director for software and network integration, my main mandate was to design, deploy, and manage the entire digital backbone of the experience, providing a solid technical foundation for the multidisciplinary team's artistic vision.

Key Responsibilities & Achievements:

Network Architecture (LAN + VLANs):
Designed and deployed a complete network infrastructure using the Ubiquiti UniFi ecosystem, optimized for low-latency critical data flow across the installation, from 3 data sites across campuses.

Experiential Pipeline:
Built the central software architecture within TouchDesigner to drive and support the installation's interactive logic and states in real-time.

Generative Fallback Architecture:
Engineered a custom non-deterministic, probabilistic democratic fallback system. During idle periods or potential sensor interruptions, this system allowed the installation's network of Nodes to autonomously and probabilistically "vote" on the environment's next state. This ensured the ecosystem remained organic, unpredictable, and "alive" without ever relying on static, pre-programmed backup loops.

Custom Tool Development:
Created specialized software solutions, including the local deployment of a Large Language Model (LLM) and the programming of a synchronized time management system.

Reliability & Quality Assurance:
Directed technical stress-test sessions to guarantee seamless stability and flawless execution during the public exhibition.

Technologies & Protocols:
Software: TouchDesigner, Local LLM deployment, Unreal Engine 5, MadMapper + MadLaser

Hardware: Ubiquiti UniFi ecosystem

Protocols: OSC, NDI, Dante

Stack: TouchDesigner, Unifi, OSC, NDI

M.O.N.A

2023 · Interactive Research · AI Interaction Designer

Mechanical Omnipresent Network Analyser

MONA is an interactive installation exploring the persistence of digital memory and surveillance through a brutalist cluster of CRT monitors. The piece acts as an omnipresent observer, greeting visitors with dark, glitch-heavy visuals driven by real-time facial analysis.

The "Crankshaft" of the Engine :
At the core of the installation lies a custom Python-based recognition pipeline that serves as the mechanical crankshaft for the visual engine. Bridging legacy hardware with modern AI, the system processes webcam streams via TouchDesigner to analyze visitors against a massive dataset in real-time.

Technical Architecture & Logic
To create a "living" archive of the exhibition, I implemented a dual-index strategy using FAISS and the NVIDIA FFHQ 1024 dataset (75k images):

First Encounter (The Doppelgänger):
For new visitors, the system vectorizes their face and queries it against the static FFHQ dataset. The result projects the user alongside their five closest matches—digital lookalikes pulled from the latent space.

Recursive Memory:
Simultaneously, the system writes the visitor's vector data to a dynamic, persistent FAISS index.

Recognition: Upon a second visit, MONA identifies the subject against this accumulated memory. It triggers a specific "Person Recognized" visual loop, confronting the viewer not just with the dataset, but with their own previous data trace.

Made with :

Clément Boucher (Project Owner and Artistic Director)

Sasha Bédard (AI Interaction Designer)

Raton Gosselin (TouchDesigner Integration)

Maxime Simard (Visual Artist)

Jean-Christophe Zephir (Sound Designer and Audio Programmer)

Dominic Roberts (Scenography)

Stack: TouchDesigner, CUDA, ArcFaces, FAISS, SQLite

Le Culte

2022 · Performance · Technical Director

Overview
As a member of Le Culte's interactive experience committee, I took on as many roles as the collective needed, from the show's backend to its portraits.

My Role: Technical Direction – Backend & Integration
For Déluge, I built the TouchDesigner patch that ran the installation: multi-video playback, light and sound control, Kinect interaction, OSC and NDI for the VJ's visuals, MIDI inputs, an idle title screen, and the integration of every other patch into one system.

It was my first patch that had to be fool-proof and fail-safe in front of an audience. Engine COMPs spread the load across CPU threads; the patch held an average of 60 FPS under full load on a consumer laptop.

Architecture: a restaurant kitchen
Kitchen: incoming data is received, cooked and parsed. Its output goes either to the Réchaud or to the Passe.
Réchaud: external inputs (MIDI, OSC) are received and parsed, then handed to the Passe.
Passe: where cooks hand plates to the floor. Python routes each cooked stream to the right receiver, starting some events and ending others.

The split keeps parsing, live input and dispatch apart, so each part can be swapped or reused in another show without touching the others.

Portraits
Over one weekend I shot portraits of more than 60 people. Lightroom Classic's tethered live view saved every frame to the right folder as it was taken, let the sitters see their image on the spot, and made one batch edit give the whole series the same base.

Stack: Resolume, TouchDesigner, Lightroom, Photoshop